The purpose of this report is to document quality-control and data transformation procedures followed to obtain methylation values (\(\beta\) and M) for the adolescent vaping study. Much of the code for this report had already been produced by the previous analyst, Cuining Liu. Further steps were taken to dissect the data-pipeline and to support any decisions made moving forward.
Methylation samples were taken using the 850K ‘EPIC’ array. All
quality control steps were conducted using the package
SeSAMe ver. 1.15.7. The steps below are not necessarily
presented in chronological order which they were taken. The order of
pre-processing steps follows the openSesame data pipeline
available here
(Zhou et al. 2018).
Samples will be evaluated for a low mean-intensity to identify potential outliers with low-quality methylation values. Outliers will be defined by samples in the bottom 1%.
The SeSAMe package uses probe ‘masking’ instead of probe
removal. These terms are interchangeable; however, it should be noted
that ‘masked’ probes are not actually removed from the dataset. They are
only ignored by SeSAMe when completing down-stream
analyses. Probes are masked in two ways:
Experiment-independent probe masking is set by a pre-determined list of probes specific to the ‘EPIC’ array. Experiment -dependent probe masking will be determined by a p-value threshold. CpG probes with a detection p-value > 0.05 in at least 10% of samples will be masked for the purpose of this study.
There are two data transformations within the sesame
pipeline. The first is a dye bias correction. Dye bias refers to bias in
the methylation values due to the performance of the red and green dyes
that create flourescence and are then interpreted as methylation values.
There are several corrections that can be applied to make the red and
green values more comparable. sesame specifically
implements a non-linear dye bias correction.
The next transformation within the openSesame pipeline
is background subtraction. The purpose of background subtraction is to
align the distribution of beta values for Infinium I and II probes in
order to make them more comparable. The default method for
sesame is a normal-exponential deconvolution using
out-of-band probes or ‘noob’ (Triche et al.
2013).
The product of the SeSAMe data pipeline is a matrix of
beta values. Beta values will be converted to M-values using the logit
transformation
\[M = log_2(\frac{\beta}{1 - \beta})\]
Visualizations will ensure the proper distribution of \(\beta\) and M-values, respectively.
Visualizations of clustering by sex, recruitment center, and vape status will help to detect any technical effects that need to be accounted for in downstream analyses.
Of the 51 subjects surveyed for this project, methylation data were
collected for 48. Two of the three subjects (SID 105 &
no_rna_meth[[1]][2]) for whom there was no transcript data
also had no methylation data. One sample (SID 102) included in the
RNASeq Analysis known as “Sample 12” also had missing methylation data.
This sample was the center of a sensitivity analysis for inclusion in
RNASeq analyse (see “sample12_sensitivity_report_2022_mm_dd.html”), and
it should be noted that this subject will not be available for
comparison. One sample (SID 144) lacked vape status in clinical metadata
and was removed after obtaining beta values for all subjects in order to
retain the benefits of that sample for normalization purposes. This
sample was also excluded from the RNASeq analysis. Downstream analyses
will include 47 samples.
| Subject ID | RNASeq ID | Methylation ID |
|---|---|---|
| 102 | Sample12 | No Data |
| 105 | No Data | No Data |
| 137 | No Data | No Data |
127 falls in the bottom 1% of mean signal intensity readings.
Considering the small sample size in this experiment, it is advisable not to remove 127, but it is worth noting for downstream analyses.
105,454 CpG sites were removed by non-experimental probe masking and 14,349 were removed by experimental probe masking (detection p-value > 0.05 in \(\ge\) 10% of samples). 119,803 probes were removed in total.
Figure 2 demonstrates the non-linear dye bias correction for a small subset of samples.
Overall, the correction worked as expected, but it should be noted that there is still a slight bias towards the green signal for higher-intensity values.
Figure 3 demonstrates the ‘noob’ method of background subtraction for a single sample.
‘noob’ background subtraction shifted the distributions of probes to have modes closer to 0 and 1 (the expected distribution of beta values). Additionally, it appears that some noise was removed from the unmethylated beta values.
Overall, \(\beta\) and M-values followed their expected distributions with peaks near 0 (unmethylated sites) and 1 (methylated sites) in the beta distribution.
Figure 4 indicates some noise around \(\beta\) = 0.5 or \(M\) = 0. This noise could indicate probes that were not removed by experiment-dependent probe masking. Noise may be reduced at a more stringent threshold for detection p-values (e.g. < 0.01).
Visualization of the samples by median X and Y chromosome intensities helps to identify samples with poor quality or samples whose clinical sex do not match predicted sex based on these values. Figure 5 displays plots of median intensity for X and Y chromosomes color-coded for both clinical and predicted sex.
Figure 5 demonstrates sample clustering by sex for median X and Y intensities. Clinical sex matches predicted sex for all samples. It should be noted that SID 122 recorded ‘non-binary’ as clinical sex, and sex was inferred by methylation data for other portions of this project.
MDS plots were made using the package minfi ver. 1.43.0
using \(\beta\) and M-values
preprocessed using the sesame pipeline. For each feature of
interest, plots were made a) using probes from the whole genome and b)
using only autosomal probes. Using only autosomal probes removes the
inherrent separation by sex to get a better idea of clustering
patterns.
The review of sample quality and clustering patterns returned one subject (SID = 127) that fell in the 1st percentile for mean intensity. There is also a clear sex effect in the data that will be accounted for by including sex as a model covariate.